Access Type

Open Access Dissertation

Date of Award

January 2011

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Robert G. Reynolds

Abstract

Evolutionary algorithms, including the Cultural Algorithms and other bio-inspired approaches are frequently used to solve problems that are not tractable for traditional approaches. Previously, research in the field of evolutionary optimization has focused on single-objective problems. On the contrary, most real-world problems involve more than one objective where these objectives may conflict with each other.

The newest implementation of the Cultural Algorithms to solve multi-objective optimization is named MOCAT. It is not the first time that the Cultural Algorithms have been used to solve multi-objective problems. Nonetheless, it is the first time that the Cultural Algorithms systematically merge techniques that have been popular in other evolutionary algorithms, such as non-domination sorting and spacing metrics, among other features. The goal of the thesis is to test whether MOCAT can efficiently handle multi-objective optimization. In addition to that, we want to observe how the knowledge sources and agent topologies within a Cultural Algorithm interact with each other during the problem solving process.

The MOCA system was evaluated against the ZDT test set proposed by Zitzler (2000). Some basic results that were produced are as follows:

1. The MOCAT system was very effective in the generation of an appropriate configuration for solving problems with different combinations of these features. Even for a given problem, as information was added to the knowledge sources, adjustments in the topologies could be made effectively.

2. As the complexity of the problems increased in terms of the number of problem features, the MOCAT system's relative performance increased.

3. A problem with just a single problem feature, such as ZDT1 and ZDT5, was often effectively solved by just using one metric guide the solution process. However, if there were multiple problems, combining the two metrics together produced a synergy that outperformed each single metric based system.

4. This synergy resulted from the fact that they rewarded spread production in different ways. The spread metric focused on global distribution while the hyper-volume tended to support local optimization.

5. The configuration of the top performing MOCAT system varied markedly from one problem to the next.

Our experiments proved the potential of applying the Cultural Algorithms on multi-objective problems and open a gate to observing internal behaviors of various knowledge sources and social fabrics.

Share

COinS